Talk:Diamond/Version History

ELO inaccuracy

About "Note: Despite lower ELO, was about .3% APS better than 1.0.", that's not surprising to me at all. Glicko-2 seems to be far more true to a full-pairing APS than ELO was too. Things like this make me glad the new server doesn't just show ELO like the old one :) --Rednaxela 22:02, 17 May 2009 (UTC)

Although I had second thoughts about setting the APS as standard ranking decisor, I must agree that ELO is not as reliable as it was on the old server(s). Mind you that ELO is calculated slightly different on this server than on the old ones. --GrubbmGait 22:11, 17 May 2009 (UTC)

ELO scores have recently taken a nosedive, for several reasons. There has been a lot of new activity recently with lots of bots being updated and a few new ones added in -- that tends to shake things up. Also, several long-running bots were removed within a short time period, notably pederson.Moron which once anchored the bottom end of the scale. It's safer to compare APS instead of ELO, especially right now. --Darkcanuck 15:18, 18 May 2009 (UTC)

I only pulled Moron because it seemed fairly pointless. I don't mind if he is returned to shore up a ratings slide, but that seems like giving a cancer patient a Band-Aid. On a related note, a long time ago I got the notion that the average ELO rating of the old rating system was 1600. I dropped the ratings list in Excel and confirmed that the ratings averaged to about 1600. At the time, most people were wondering what I was smoking, dismissing 1600 as anything of relevance. I recently did another averaging of the ratings and found ELO to average at 1413 and Glicko-2 averages 1608. Dunno if it really means anything, but ELO certainly doesn't compare to the old ratings.--Martin 16:24, 18 May 2009 (UTC)

Avoiding recent enemy locations

Well, I found a bug in my risk calculation for avoiding recent enemy locations (fixed but not tuned in 1.071). I really feel like this must be a good idea (because bullets are likely to be headed to those spots), but I hadn't found any rating boost from it yet. Hopefully I can find some points in a re-tuned, bug-free version of this... --Voidious 17:25, 21 May 2009 (UTC)

1.11* bugs / fixes

Man, this Performance Enhancing Bug really drove me nuts, especially since I have so much trouble testing such a little thing. I think I can live with a .15% APS drop in my "bug-free" 1.115 (versus the "buggy" 1.111). I'm sure I'll continue tuning the randomness of his movement in the future, anyway, and I know there are more performance gains to be had elsewhere. This is long and boring, but I feel compelled to write it out, even if just to have the info "out there" somewhere.

The bug was with my random direction change timer. This timer influences the bot to reverse direction at a random interval (with a risk added for disobeying the timer). If the timer trips while the bot is moving with negative velocity, there would be a risk associated with continuing in that direction, and the bot would soon change direction towards positive velocity. Before the timer trips, there's a risk associated with reversing direction.

With the bug, once the velocity went from negative to zero (or positive), it mixed up its directions and thought that negative velocities were the safer direction. This was caused by two things: it always considered zero velocity to be the same as positive, and it was comparing the possible future heading to the heading from one tick ago (instead of heading from the current tick). So here's the problem scenario:

• Diamond is moving with negative velocity.
• The timer trips, so now there's a risk in continuing in this direction.
• He begins to change direction, eventually hitting zero velocity (or, rarely, as high as +1).
• He notices he's changed direction (because zero is the other direction) and resets the timer, meaning now there is a risk with changing direction.
• But he compares possible movement direction against that of one tick ago, which is a negative velocity, so he thinks that way is the "same" direction.

In short, the direction changing worked fine when moving with positive velocity, but for negative velocity, it was more likely to just stop and then continue in that direction. At this point, I've just removed this timer and tuned the other randomizing factor (risk from recent locations) a bit more.

--Voidious 15:32, 26 May 2009 (UTC)

One thing at a time

The changes in 1.12 and 1.121 are a great example of why you should follow the "one change at a time" dogma. From 1.072 (best version at the time) to 1.08, I removed one thing: risk from recent enemy locations, and added another: risk factor based on damage given to enemy. 1.08 went down 0.1% APS, so I thought, "well that's barely beyond the margin of error, I'll just leave it". From 1.115 to 1.12, I restored the risk to recent enemy locations and saw a .25% APS drop. From 1.115 to 1.121, I removed the damage given risk factor and am seeing a .4% APS gain (pairings almost complete). Yay! --Voidious 18:25, 27 May 2009 (UTC)

Bah, a few more battles and it's not even above 1.115. I guess I should be patient (still only at 1,000 battles). Oh well... =) --Voidious 21:14, 27 May 2009 (UTC)

One-on-one

At least you start working with One-on-one! » Nat | Talk » 14:49, 1 June 2009 (UTC)

Just a little bit to keep myself sane. =) Melee is really hard. Spending a day on 1v1 is therapeutic because I actually know what I'm doing. =) --Voidious 14:56, 1 June 2009 (UTC)

Also, I know I've covered this topic ad nauseum already, but I just have to vent: flatteners are so evil!! They entice you with their 50+ scores against CC in the MC2K6 and then they destroy your RoboRumble rating... =) --Voidious 16:02, 1 June 2009 (UTC)

Evil?!? Not if going for PL ;) --Rednaxela 18:52, 1 June 2009 (UTC)

Yeah, yeah, that's what the devilish flattener always says! =) Just kidding, obviously I'll try to tune it better. But they sure are touchy buggers. --Voidious 19:13, 1 June 2009 (UTC)

Once you get them tuned they can actually give you points. I know on DrussGT the flattener gives a good 5 ELO points at least. The trick is putting your bot against the enemy with the lowest hitrate that you would want the flattener enabled against (Ascendant comes to mind) and then putting the threshold just a tiny bit below that. --Skilgannon 19:56, 1 June 2009 (UTC)

Yeah, I've tested removing the flattener from Dookious and it costs me points (I think it was about 5, too, not to mention killing PL score). The thresholds there are really well tuned and conservative. I'm just feeling greedy now. =) I think I'm going to try to come up with something more clever than just hit percentages (even my carefully normalized ones), like measuring the adaptation rate of the enemy gun, for enabling criteria. Turning the flattener back off when the enemy hit-% drops below the threshold (as I do in Dookious) seems silly from one perspective, because it could just mean the flattener's working, but I've always felt that permanently enabling it carried too much risk. --Voidious 20:05, 1 June 2009 (UTC)

Hmm... Normalized hitrates... I should compare those used in RougeDC's firepower selection with the ones you use for flattener enabling :) --Rednaxela 01:10, 2 June 2009 (UTC)

I tally bullets fired as normal, but bullet hits are weighted by distance and precise escape angle range. So a bullet hit from twice as far away or in a situation with twice as big a max escape angle would count as two hits. (Edit: Oh yeah, I also subtract 1 from bullets fired for onBulletHitBullet.) For my bullet power management experiment in Dookious 1.60, I normalized out the same stuff in a different way, but I didn't use precise max escape angles, because I was feeding the escape angle into a formula for testing 300 bullet powers each tick. =) --Voidious 01:31, 2 June 2009 (UTC)

I've also been thinking about ways to make the flattener smarter, and one was by enabling it when entropy shows that it is predicting where they will shoot better than the regular hit surfing. I haven't tried it yet though. --Skilgannon 20:33, 1 June 2009 (UTC)

Wow, nearly 1% APS improvement with Dooki's gun, meaning more than half the difference between Diamond and Dookious lies in the gun. I guess the movement is already capable of some pretty nice PL performance, too. Exciting! --Voidious 13:59, 13 July 2009 (UTC)

MeleeRumble cruelty

So I feel 99% certain that 1.196 is functionally equivalent to 1.183, but that's a big discrepancy: 1.196 vs 1.183. I recompiled the source I had for 1.183 for a re-release, and it's coming in somewhere between: 1.183b vs 1.183. A binary comparison of the .class files shows that they are all the same besides one which I never update, so I'm confident the source is right. The MeleeRumble is a cruel, frustrating beast... --Voidious 16:02, 10 July 2009 (UTC)

It does make a big difference score to one robot if you fight it under a set of sample robot (which exist in melee) and a battle full of ABC's, rozu's and justin's robots. Although the difference you point isn't as much as I expected. Perhaps we should have a better way to control the melee score. Perhaps we need to weight the score base on the opponent level (which can take from the ranking). But it's a work. » Nat | Talk » 17:29, 10 July 2009 (UTC)

Multiple gun waves experiment

Well, I'm really surprised this worked, and now I'm wondering if anyone else has tried it before. Whenever I fire a gun wave in melee, I now fire an additional gun wave from the last known location of each (still living) enemy (besides the target). The idea is that I can easily collect gun data from the perspectives of all bots on the battle field, and that hopefully, I'll get a better/faster picture of the enemy movements this way.

A couple bug fixes and a bit of polishing later, I had a 0.74% improvement in my test bed from the new gun waves, on top of a 0.14% improvement with the tweak to the "number of bots" weight (also in 1.283). I'm now looking at a ~0.6% APS improvement in the rumble for 1.284 over 1.282, and not too far behind Aleph. Yay!

--Voidious 15:35, 24 August 2009 (UTC)

I've thought about something similar before, but I've yet to build a real melee gun. Really nice stuff! What I wonder, is what the results would be like if you added other points like the corners to that list, since many bots move relative to the corners as well... --Rednaxela 15:43, 24 August 2009 (UTC)

Well, with the way my gun records and reconstructs firing angles, firing waves from the corners for that reason wouldn't have the same effect as if you were using GuessFactors. (The enemy movement isn't recorded relative to the wave source, but relative to the enemy's initial heading.) But indeed, firing waves from additional sources is a good idea to try, and you just gave me the idea to try a "heading relative to nearest corner" attribute... --Voidious 16:03, 24 August 2009 (UTC)

Congrats, that's a huge and clever improvement! --Positive 15:58, 24 August 2009 (UTC)

Thanks! Still a ways to go before I catch Aleph or Portia 1.13, and I don't even want to think about Shadow 3.83, but I'm very happy for now. =) --Voidious 16:03, 24 August 2009 (UTC)

Very clever indeed! That kind of specific melee gun idea is exactly what made Shadow finally break away from Aleph in the rumble. You (and Positive) are not as far from Shadow as you might think, imo. --ABC 16:58, 24 August 2009 (UTC)

MeleeRumble 2nd place

It seems that no-one noticed, due to the ....Hawk hype, that Diamond passed Aleph and reached second place. Congratulations Voidious! And do I read the results of DiamondHawk correctly if I state that your movement is better than Shadows, but your gun is holding you back?? --GrubbmGait 05:38, 28 August 2009 (UTC)

Thanks for noticing. =) However, Portia 1.13 actually still has 2nd place by a small margin if he were un-retired. I'm still pretty excited about passing Aleph, though! And yes, I'm surprised by how many points I have left in my gun (and working on it now =)). If the *Hawk results translate linearly, this would put Diamond's movement ahead of Shadow 3.84, which is really exciting (even to be close!), but still a little behind 3.83. --Voidious 05:44, 28 August 2009 (UTC)

How about ask ABC to create Shadond? I'm curious if he will get to the throne (or even SHA3.83?) » Nat | Talk » 13:47, 28 August 2009 (UTC)

Now you know what I meant by you and Positive being closer to Shadow than you thought ;). Shadow 3.83 melee movement is just a normal (and very old) Minimum Risk movement, the melee strategy page on this wiki has long contained all the tricks I use. Aleph's movement has probably always been slightly better than Shadow's, but my gun made the difference. At this time I suspect that the strongest combination would be my gun with Portia's movement. --ABC 15:00, 28 August 2009 (UTC)

Well, perhaps the very strongest combination would be... Shadow's gun, Portia's Melee movement, and Diamond's 1v1 movement, but the 1v1 part would only make a small difference I'm sure... :) --Rednaxela 15:04, 28 August 2009 (UTC)

Nah, DrussGT's movement is stronger than Diamond's 1v1. » Nat | Talk » 15:16, 28 August 2009 (UTC)

ShadowHawk and DiamondHawk gave me a pretty good idea where my gun and movement stand, so I don't feel a need to request Shadond. (That is a hilarious name, though.) I'm quite content to wait until I can challenge Shadow myself. =) ABC (or Positive) is free to make a hybrid if he likes - my code is very pluggable, as always. 1.30 is officially at #2 now, so maybe I'll focus on 1v1 for a bit. =) --Voidious 15:21, 28 August 2009 (UTC)

The only obstacle is RWPCL. So unless you officially gave them permissions, they can't =) » Nat | Talk » 15:32, 28 August 2009 (UTC)

Oh bah, the statement "ABC (or Positive) is free to" is good enough for all intents and purposes :) --Rednaxela 15:47, 28 August 2009 (UTC)

(Edit conflict) Diamond 1.30 has definitely passed Portia 1.13 now, so congrats on reaching 2nd place! It's interesting to think about what combination would be best. I think Shadow's movement and gun are quite good for the first few turns of a 1v1 fight, because they don't wait for waves to reach the target (something GF-targeting does have to wait for). In melee, you often only get a few shots while down to 1v1, so it's extra important to have a quick start. --Positive 15:27, 28 August 2009 (UTC)

Thanks! I don't understand, why would GF targeting have to wait for waves to reach the target before firing? Diamond sometimes stays still for a moment when it gets down to 1v1, but I'm not really sure why, actually... I should investigate that. --Voidious 19:51, 28 August 2009 (UTC)

What I mean is that GF-guns can only use a simple targeting method to aim before they have any data, and they only have enough data to make an informed shot after a few waves have hit. A pattern matching can already detect and simulate movement (for example, stop 'n go) before its first waves hit. --Positive 21:45, 28 August 2009 (UTC)

Oh, of course. But in a Melee battle, you can still use the targeting data you collected during Melee when it gets down to 1v1, right? I don't see why a GF gun can't use that Melee data (as I'm sure Shadow does). Shadow is an awesomely quick learner, though, there's no doubt about that. --Voidious 21:56, 28 August 2009 (UTC)

That's true, you could use a combination of the melee data and new data. However, I think most advanced robots behave or will behave differently from melee in 1v1, so it will be or is better to use new data as fast as possible. Also, if you look at Shadow's debugging graphics, it seems like it doesn't use the melee data. :P --Positive 22:06, 28 August 2009 (UTC)

Really? Interesting. Diamond uses "number of bots" as one of its gun dimensions, so 1v1 data is favored, but it uses whatever it has. If I were using GF + VCS, I would probably sum a lower-weighted buffer that doesn't segment on # of bots and a higher weighted one that does, so it would have something to work from before collecting the 1v1 waves. =) --Voidious 22:11, 28 August 2009 (UTC)

Oh, very nice - congrats! Perhaps this will trigger the release of Shadow 3.85? --Darkcanuck 15:56, 28 August 2009 (UTC)

I sure hope not... =) --Voidious 19:51, 28 August 2009 (UTC)

Wow, 1.31 is 0.4 APS from Shadow 3.84, congrats! Now I really have to see if I can squeeze some more points from Shadow's gun/movement. --ABC 12:30, 29 August 2009 (UTC)

Thanks! =) Now that I've got it working, your Melee Gun design is so powerful, it almost feels dirty. Thanks for sharing. I'll probably focus on 1v1 for a bit now, anyway. And seeing as rating points are so much harder to find as you ascend towards #1, I think 3.83 still has a pretty solid lead over Diamond. --Voidious 17:09, 29 August 2009 (UTC)

I just noticed, you are now second in both melee and 1v1 =) Great work you've been doing on Diamond. I'm curious, how much do your melee and 1v1 code interact? For instance, at the end of a melee when there are only 2 bots left (1 being Diamond), does your 1v1 movement kick in with no values? Or does it have some sort of idea where the enemy will be aiming due to it's observations during the melee? --Skilgannon 17:59, 29 August 2009 (UTC)

Thanks! While there's nothing quite like being #1 ;), I'm really satisfied with Diamond's progress. It's also great fun having active competition from Positive, ABC, and Justin... Hopefully, I can give you that "pleasure" on the 1v1 side sometime. =) I know it gets lonely up there. As to your question:

• On the movement side, they are totally separate, much like you said. When it gets to 1v1, the surfing kicks in with preloaded HOT-avoidance, not knowing anything from melee bullet hits. Of course it saves surf stats per-enemy across rounds, too. If I had some Portia-style melee bullet dodging, I'd surely try to integrate it more, but as of now, Diamond's melee movement doesn't even notice energy drop.
• On the gun side, they are more integrated. One of the gun dimensions is number of bots alive, but it doesn't discard melee data for 1v1 scenarios or vice versa. Wall distance is calculated differently (though recorded in the same slot), I use different weights when aiming (including some dimensions being used in only 1v1 or melee), and now the melee gun uses the Shadow/Melee Gun technique.

--Voidious 19:10, 29 August 2009 (UTC)

Hmm. So the gun works pretty optimally between melee/1v1 but the movement doesn't. The movement I've been working on expands seamlessly from 1v1 to melee, with full surfing capabilities, and full learning based not only on hits to yourself, but to all other bots that you can scan. The logic is amazingly complex, with decisions about bullet hits needing to be deferred until more data is collected and checked for bullet bonuses etc, but I see no reason (aside from CPU limits) why it isn't perfectly viable to learn how everybody is shooting all at once =) This is the main reason I haven't progressed that far with it - it's such a massive project that I'm not sure where to begin, and I'm also scared that I'll spend several hundred hours in development and then the whole thing will be a flop =) --Skilgannon 20:35, 29 August 2009 (UTC)

Sounds very cool, and very ambitious. =) I see how a lot can be approximated about enemy targeting during melee, I'm just very skeptical it can be precise enough to be useful for Wave Surfing, which thrives on ultimate precision. But Portia has found a lot of success, ABC is trying likewise, and it sounds like Rednaxela has some good melee dodging now too, so maybe it's worth a shot. Good luck, in any case - it'd be way cool to see you enter the melee arena. --Voidious 21:30, 29 August 2009 (UTC)

I just saw that Diamond 1.31 is PL king at the moment (100% Pairs won), congrats! --Positive 00:14, 1 September 2009 (UTC)

Thanks... Though the top bots are so always so close and I release so many versions, it was bound to happen eventually. =) Likewise, great work with Portia 1.19 - great way to start the semester, eh? Cheers, --Voidious 02:30, 1 September 2009 (UTC)

Performance Enhancing Bug in 1.32

Hm, well it seems that the bug fixed in 1.323 was slightly performance enhancing, and thusly says to me that something isn't ideal about using inverse-distance weighting for scans... Hmm... --Rednaxela 14:58, 31 August 2009 (UTC)

Yeah, I don't know what's up with that. It's not that far beyond the margin of error, but still, I was hoping for a boost. I'm rewriting a lot of my data logging / danger projection into a bigger / badder system, anyway, so I actually don't care all that much. (It was actually while writing up my new system that I noticed this bug.) For what it's worth, in my experience, weighting scans by inverse distance is quite essential against simple targeters, where your most similar scans tell you exactly where the danger is and the rest are just noise. --Voidious 15:30, 31 August 2009 (UTC)

Well, I've found in various experiments over time that weighting scans some function of distance is essential indeed, but that a simple inverse is suboptimal. --Rednaxela 16:07, 31 August 2009 (UTC)

What, according to your experiments would be closer to optimal then? The inverse of some power of the distance perhaps?--Navajo 22:01, 1 September 2009 (UTC)

I'm also curious to hear what Rednaxela thinks... I used to use inverse square root of distance in Lukious, and I think Skilgannon tends to prefer inverse manhattan distance. Diamond is one of the strongest DC movements and is currently using a simple division by distance, for whatever that's worth. (In terms of rumble points, I think it just passed Hydra for the strongest DC movement.)

Semi-off-topic ramblings: I think we end up tuning around a lot of arbitrary things in our very complex bots, so when we change something and lose points, we think the change was universally "bad", when it really was just bad in our specific case. That's just a hunch, partially based on how often I and others tweak values that we've never tweaked and (seem to) find that our first instinct for that value cannot be improved upon.

--Voidious 23:00, 1 September 2009 (UTC)

(OT) I was just thinking the same -- 99% of my tweaks end up with similar or slightly worse performance, it's usually the bigger changes that yield results. --Darkcanuck 03:41, 2 September 2009 (UTC)

Yes, sometimes I make a change that when tested with several seasons yields some APS gain and I never upload a version of YersiniaPestis (after 1.3.7) that is not able to beat all my test bed bots on average, but on the rumble luck plays a huge role. Right now version 3.0 loses to Locke, and I haven't tested against him, but I'm most confident that if they fight some more times that would change, the only reason I haven't really cared about that is because Shadow is currently losing to two bots :). --zyx 04:09, 2 September 2009 (UTC)

Well, I had the most success with things like 1/distance^n or 1/n^distance type things for some magic tweaked value of n. What I found most interesting of all though, was that the best values were highly specific to the dimensions and surfing algorithm in question. When I said that I found simple inverse to be suboptimal, I didn't really have anything very specific in mind that I expected to be better for Diamond, but I had a doubt reinforced by this PerformanceEnhancingBug, that it was quite unlikely that a plain inverse was being truly optimal --Rednaxela 05:59, 2 September 2009 (UTC)